Analyze and optimize your AI features.
Log requests, add caching, and run experiments. Just two lines of code to get started.
How it works
Add 2 lines of code to get started
Warehouse every OpenAI and Anthropic request to your PostgreSQL database. Use logs to analyze, evaluate, and generate datasets.
Analyze usage & run experiments
We store a customizable JSON object so you can granularly monitor usage, calculate cost, run evaluations, and fine-tune models.
Optimize with caching & batching
Enable caching to reduce costs and latency. Get full transparency into OpenAI's Batch and Files APIs using our built-in proxy support.
AI gateway designed for engineers
"We experiment with LLM models, settings, and optimizations. Velvet made it easy to implement logging and caching. And we're storing training sets to eventually fine-tune our own models.
"Velvet gives us a source of truth for what's happening between the Revo copilot, and the LLMs it orchestrates. We have the data we need to run evaluations, calculate costs, and quickly resolve issues."
"Our engineers use Velvet daily. It monitors AI features in production, even opaque APIs like batch. The caching feature reduces costs significantly. And, we use the logs to observe, test, and fine-tune."
Flexible infrastructure for scale
Warehouse logs
Log every request to your database. Include custom metadata.
Analyze usage & costs
We store data as a JSON object so you can granularly query logs.
Cache requests
Return identical results in milliseconds. Reduce costs and latency.
Batch jobs
Unlock extra data so you can query each file inside the batch.
Evaluate & experiment
Utilize datasets for model evaluations, prompt testing, and fine-tuning.
AI SQL editor
Use an editor that automates SQL on top of large datasets.
Q & A
Articles from Velvet
Why Find AI logs OpenAI requests with Velvet
AI-powered B2B search engine logged 1,500 requests per second.
Create a fine-tuning dataset for gpt-4o-mini
Use Velvet to identify and export a fine-tuning dataset.
Cache LLM requests to reduce latency and costs
Return results in milliseconds and don't waste calls on identical requests.